Addressing the challenge of distinguishing between single and multiple sensor failures in nuclear power plants under varying operating conditions due to their highly similar manifestations, this paper proposes a distinction model for single and multiple sensor failures based on Principal Component Analysis (PCA) and improved Long Short-Term Memory (LSTM) networks within a cyclic monitoring mechanism. Firstly, PCA is utilized for data dimensionality reduction and feature extraction. Combined with squared prediction error statistics and contribution rates, this process achieves anomaly monitoring and failed sensor localization. Secondly, LSTM network fused with a convolutional autoencoder reconstructs data for failed sensors. Finally, monitoring and reconstruction models are synergistically integrated through a cyclic monitoring mechanism, where reconstructed data replaces failed measurements for iterative re-monitoring. The distinction between single and multiple sensor failures is achieved through predetermined cyclic iterations. Validation using data obtained from a nuclear power plant simulator under varying operating conditions demonstrates the model’s effectiveness. The results demonstrate that the reconstruction model achieves a mean absolute percentage error of 4.9%, while the integrated model attains a discrimination accuracy of 97.5% for distinguishing single and multi-sensor failures. This model could reduce operator stress to some extent and reduce nuclear power plant accidents caused by human error.